def partial_trace_wf_keep_first_cupy(iwf: cupy.ndarray, n):
# TODO: improve the cuda version of partial_trace_wf.
# For example, cleverly adjust the blockDim to deal with the other case
assert iwf.flags.c_contiguous
assert iwf.dtype == default_dtype
iwf_conj = iwf.conj()
nqb = int(math_log2(iwf.shape[0]))
m = nqb - n
m_idx = 2 ** m
n_idx = 2 ** n
rho = zeros(shape=(n_idx, n_idx), dtype=default_dtype, order="C")
# Here we simply use the threadDim for i, j in the cuda code.
threads_per_bloch = 32
threadDim = (threads_per_bloch, threads_per_bloch)
x = (n_idx + (threads_per_bloch - 1)) // threads_per_bloch
blockDim = (x, x)
partial_trace_wf_keep_first_cuda(
grid=blockDim,
block=threadDim,
args=(
iwf,
iwf_conj,
rho,
m,
m_idx,
n_idx,
),
)
return rho
def partial_trace_wf_cupy(iwf: cupy.ndarray, retain_qubits):
nqb = int(math_log2(iwf.shape[0]))
if len(retain_qubits) == nqb:
return outer(iwf, iwf.conj())
iwf = iwf.reshape([2] * nqb, order="C")
retain_qubits = sorted(retain_qubits)
for idx in range(len(retain_qubits)):
r = retain_qubits[idx]
if idx != r:
iwf = iwf.swapaxes(idx, r)
iwf = iwf.reshape((2 ** nqb,))
return partial_trace_wf_keep_first_cupy(iwf, len(retain_qubits))